Quantitative gait analysis is vital for understanding human movement, with vertical Ground Reaction Forces (vGRFs) being a key metric. Traditional methods to measure vGRFs, such as force plates and treadmills, are bulky, expensive, and may not reflect natural gait patterns. This study introduces a physics informed deep learning framework that uses ankle position data in the sagittal plane to predict vGRFs from marker-based motion capture. The framework includes sequence-to-sequence models such as LSTM, GRU, and one-dimensional convolutional layers with LSTM, using both data-driven (Seq2Seq) and physics-based (Physics-Informed Neural Network, PiNN). Internal dataset results showed that Seq2Seq outperformed PiNN on the left limb, while PiNN showed better performance on the right limb in some configurations. Both models experienced increased RMSE on external datasets, highlighting challenges in generalization. Spearman correlation analysis indicated PiNN had a stronger relationship with measured vGRFs in the internal dataset, but Seq2Seq demonstrated better generalization. RMSE values ranged from 11.55% to 16.00% body weight (BW) for Seq2Seq models and 11.55% to 17.85% BW for PiNN. Although Seq2Seq models had greater variability, they closely matched measured values in certain configurations. These findings suggest PiNN models effectively capture vGRF patterns, but further refinement and training on diverse datasets are needed to improve generalization. Future research should focus on incorporating additional physical constraints to enhance model robustness in real-world scenarios.